Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,221 +1,37 @@
|
|
1 |
-
from
|
2 |
-
import
|
3 |
-
import
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
|
39 |
-
sum_embeddings = torch.sum(token_embeddings * input_mask_expanded, 1)
|
40 |
-
sum_mask = torch.clamp(input_mask_expanded.sum(1), min=1e-9)
|
41 |
-
return (sum_embeddings / sum_mask).squeeze(0)
|
42 |
-
|
43 |
-
def average_embedding(embedding_list):
|
44 |
-
tensors = torch.stack([emb for _, emb in embedding_list])
|
45 |
-
return tensors.mean(dim=0)
|
46 |
-
|
47 |
-
def get_cosine_similarity(vec1, vec2):
|
48 |
-
if torch.is_tensor(vec1): vec1 = vec1.detach().numpy()
|
49 |
-
if torch.is_tensor(vec2): vec2 = vec2.detach().numpy()
|
50 |
-
vec1 = vec1.flatten()
|
51 |
-
vec2 = vec2.flatten()
|
52 |
-
dot_product = np.dot(vec1, vec2)
|
53 |
-
norm_a = np.linalg.norm(vec1)
|
54 |
-
norm_b = np.linalg.norm(vec2)
|
55 |
-
if norm_a == 0 or norm_b == 0: return 0
|
56 |
-
return dot_product / (norm_a * norm_b)
|
57 |
-
|
58 |
-
def get_combined_scores(query_vector, embedding_list, all_good_embeddings, avg_weight=0.6):
|
59 |
-
results = []
|
60 |
-
for name, emb in embedding_list:
|
61 |
-
avg_similarity = get_cosine_similarity(query_vector, emb)
|
62 |
-
individual_similarities = [get_cosine_similarity(good_emb, emb) for _, good_emb in all_good_embeddings]
|
63 |
-
avg_individual_similarity = sum(individual_similarities) / len(individual_similarities)
|
64 |
-
combined_score = avg_weight * avg_similarity + (1 - avg_weight) * avg_individual_similarity
|
65 |
-
results.append((name, emb, combined_score))
|
66 |
-
results.sort(key=lambda x: x[2], reverse=True)
|
67 |
-
return results
|
68 |
-
|
69 |
-
def find_best_ingredients(required_ingredients, available_ingredients, max_ingredients=6, avg_weight=0.6):
|
70 |
-
required_ingredients = list(set(required_ingredients))
|
71 |
-
available_ingredients = list(set([i for i in available_ingredients if i not in required_ingredients]))
|
72 |
-
if not required_ingredients and available_ingredients:
|
73 |
-
random_ingredient = random.choice(available_ingredients)
|
74 |
-
required_ingredients = [random_ingredient]
|
75 |
-
available_ingredients = [i for i in available_ingredients if i != random_ingredient]
|
76 |
-
if not required_ingredients or len(required_ingredients) >= max_ingredients:
|
77 |
-
return required_ingredients[:max_ingredients]
|
78 |
-
if not available_ingredients:
|
79 |
-
return required_ingredients
|
80 |
-
embed_required = [(e, get_embedding(e)) for e in required_ingredients]
|
81 |
-
embed_available = [(e, get_embedding(e)) for e in available_ingredients]
|
82 |
-
num_to_add = min(max_ingredients - len(required_ingredients), len(available_ingredients))
|
83 |
-
final_ingredients = embed_required.copy()
|
84 |
-
for _ in range(num_to_add):
|
85 |
-
avg = average_embedding(final_ingredients)
|
86 |
-
candidates = get_combined_scores(avg, embed_available, final_ingredients, avg_weight)
|
87 |
-
if not candidates: break
|
88 |
-
best_name, best_embedding, _ = candidates[0]
|
89 |
-
final_ingredients.append((best_name, best_embedding))
|
90 |
-
embed_available = [item for item in embed_available if item[0] != best_name]
|
91 |
-
return [name for name, _ in final_ingredients]
|
92 |
-
|
93 |
-
def skip_special_tokens(text, special_tokens):
|
94 |
-
for token in special_tokens: text = text.replace(token, "")
|
95 |
-
return text
|
96 |
-
|
97 |
-
def target_postprocessing(texts, special_tokens):
|
98 |
-
if not isinstance(texts, list): texts = [texts]
|
99 |
-
new_texts = []
|
100 |
-
for text in texts:
|
101 |
-
text = skip_special_tokens(text, special_tokens)
|
102 |
-
for k, v in tokens_map.items(): text = text.replace(k, v)
|
103 |
-
new_texts.append(text)
|
104 |
-
return new_texts
|
105 |
-
|
106 |
-
def validate_recipe_ingredients(recipe_ingredients, expected_ingredients, tolerance=0):
|
107 |
-
recipe_count = len([ing for ing in recipe_ingredients if ing and ing.strip()])
|
108 |
-
expected_count = len(expected_ingredients)
|
109 |
-
return abs(recipe_count - expected_count) == tolerance
|
110 |
-
|
111 |
-
def generate_recipe_with_t5(ingredients_list, max_retries=5):
|
112 |
-
original_ingredients = ingredients_list.copy()
|
113 |
-
for attempt in range(max_retries):
|
114 |
-
try:
|
115 |
-
if attempt > 0:
|
116 |
-
current_ingredients = original_ingredients.copy()
|
117 |
-
random.shuffle(current_ingredients)
|
118 |
-
else:
|
119 |
-
current_ingredients = ingredients_list
|
120 |
-
ingredients_string = ", ".join(current_ingredients)
|
121 |
-
prefix = "items: "
|
122 |
-
generation_kwargs = {
|
123 |
-
"max_length": 512, "min_length": 64, "do_sample": True,
|
124 |
-
"top_k": 60, "top_p": 0.95
|
125 |
-
}
|
126 |
-
inputs = t5_tokenizer(
|
127 |
-
prefix + ingredients_string, max_length=256, padding="max_length",
|
128 |
-
truncation=True, return_tensors="jax"
|
129 |
-
)
|
130 |
-
output_ids = t5_model.generate(
|
131 |
-
input_ids=inputs.input_ids, attention_mask=inputs.attention_mask, **generation_kwargs
|
132 |
-
)
|
133 |
-
generated = output_ids.sequences
|
134 |
-
generated_text = target_postprocessing(t5_tokenizer.batch_decode(generated, skip_special_tokens=False), special_tokens)[0]
|
135 |
-
recipe = {}
|
136 |
-
sections = generated_text.split("\n")
|
137 |
-
for section in sections:
|
138 |
-
section = section.strip()
|
139 |
-
if section.startswith("title:"):
|
140 |
-
recipe["title"] = section.replace("title:", "").strip().capitalize()
|
141 |
-
elif section.startswith("ingredients:"):
|
142 |
-
ingredients_text = section.replace("ingredients:", "").strip()
|
143 |
-
recipe["ingredients"] = [item.strip().capitalize() for item in ingredients_text.split("--") if item.strip()]
|
144 |
-
elif section.startswith("directions:"):
|
145 |
-
directions_text = section.replace("directions:", "").strip()
|
146 |
-
recipe["directions"] = [step.strip().capitalize() for step in directions_text.split("--") if step.strip()]
|
147 |
-
if "title" not in recipe:
|
148 |
-
recipe["title"] = f"Rezept mit {', '.join(current_ingredients[:3])}"
|
149 |
-
if "ingredients" not in recipe:
|
150 |
-
recipe["ingredients"] = current_ingredients
|
151 |
-
if "directions" not in recipe:
|
152 |
-
recipe["directions"] = ["Keine Anweisungen generiert"]
|
153 |
-
if validate_recipe_ingredients(recipe["ingredients"], original_ingredients):
|
154 |
-
return recipe
|
155 |
-
else:
|
156 |
-
if attempt == max_retries - 1: return recipe
|
157 |
-
except Exception as e:
|
158 |
-
if attempt == max_retries - 1:
|
159 |
-
return {
|
160 |
-
"title": f"Rezept mit {original_ingredients[0] if original_ingredients else 'Zutaten'}",
|
161 |
-
"ingredients": original_ingredients,
|
162 |
-
"directions": ["Fehler beim Generieren der Rezeptanweisungen"]
|
163 |
-
}
|
164 |
-
return {
|
165 |
-
"title": f"Rezept mit {original_ingredients[0] if original_ingredients else 'Zutaten'}",
|
166 |
-
"ingredients": original_ingredients,
|
167 |
-
"directions": ["Fehler beim Generieren der Rezeptanweisungen"]
|
168 |
-
}
|
169 |
-
|
170 |
-
# Kernlogik, die von der FastAPI-Route aufgerufen wird
|
171 |
-
def process_recipe_request_logic(required_ingredients, available_ingredients, max_ingredients, max_retries):
|
172 |
-
"""
|
173 |
-
Kernlogik zur Verarbeitung einer Rezeptgenerierungsanfrage.
|
174 |
-
"""
|
175 |
-
if not required_ingredients and not available_ingredients:
|
176 |
-
return {"error": "Keine Zutaten angegeben"}
|
177 |
-
try:
|
178 |
-
optimized_ingredients = find_best_ingredients(
|
179 |
-
required_ingredients, available_ingredients, max_ingredients
|
180 |
-
)
|
181 |
-
recipe = generate_recipe_with_t5(optimized_ingredients, max_retries)
|
182 |
-
result = {
|
183 |
-
'title': recipe['title'],
|
184 |
-
'ingredients': recipe['ingredients'],
|
185 |
-
'directions': recipe['directions'],
|
186 |
-
'used_ingredients': optimized_ingredients
|
187 |
-
}
|
188 |
-
return result
|
189 |
-
except Exception as e:
|
190 |
-
return {"error": f"Fehler bei der Rezeptgenerierung: {str(e)}"}
|
191 |
-
|
192 |
-
# --- FastAPI-Implementierung ---
|
193 |
-
app = FastAPI() # Deine FastAPI-Instanz
|
194 |
-
|
195 |
-
class RecipeRequest(BaseModel):
|
196 |
-
required_ingredients: list[str] = []
|
197 |
-
available_ingredients: list[str] = []
|
198 |
-
max_ingredients: int = 7
|
199 |
-
max_retries: int = 5
|
200 |
-
# Abwärtskompatibilität: Falls 'ingredients' als Top-Level-Feld gesendet wird
|
201 |
-
ingredients: list[str] = []
|
202 |
-
|
203 |
-
@app.post("/generate_recipe") # Der API-Endpunkt für Flutter
|
204 |
-
async def generate_recipe_api(request_data: RecipeRequest):
|
205 |
-
"""
|
206 |
-
Standard-REST-API-Endpunkt für die Flutter-App.
|
207 |
-
Nimmt direkt JSON-Daten an und gibt direkt JSON zurück.
|
208 |
-
"""
|
209 |
-
final_required_ingredients = request_data.required_ingredients
|
210 |
-
if not final_required_ingredients and request_data.ingredients:
|
211 |
-
final_required_ingredients = request_data.ingredients
|
212 |
-
|
213 |
-
result_dict = process_recipe_request_logic(
|
214 |
-
final_required_ingredients,
|
215 |
-
request_data.available_ingredients,
|
216 |
-
request_data.max_ingredients,
|
217 |
-
request_data.max_retries
|
218 |
-
)
|
219 |
-
return JSONResponse(content=result_dict)
|
220 |
-
|
221 |
-
print("INFO: FastAPI application script finished execution and defined 'app' variable.")
|
|
|
1 |
+
from fastapi import FastAPI
|
2 |
+
from fastapi.responses import JSONResponse
|
3 |
+
from pydantic import BaseModel # Bleibt, da FastAPI es für Request Body Parsing nutzt
|
4 |
+
|
5 |
+
# Deine FastAPI-Instanz
|
6 |
+
app = FastAPI(title="Minimal FastAPI Test App")
|
7 |
+
|
8 |
+
# Eine einfache Request-Modell-Klasse (auch wenn wir sie hier nicht wirklich nutzen,
|
9 |
+
# zeigt es, dass Pydantic funktioniert)
|
10 |
+
class TestRequest(BaseModel):
|
11 |
+
message: str = "Hello"
|
12 |
+
|
13 |
+
# Ein einfacher "Hello World"-Endpunkt, der auf POST-Anfragen reagiert
|
14 |
+
@app.post("/test")
|
15 |
+
async def test_api_endpoint(request_data: TestRequest):
|
16 |
+
"""
|
17 |
+
Ein einfacher Test-Endpunkt.
|
18 |
+
Gibt eine Begrüßungsnachricht zurück, die die empfangene Nachricht enthält.
|
19 |
+
"""
|
20 |
+
print(f"INFO: Received test request with message: {request_data.message}") # Log für den Space
|
21 |
+
return JSONResponse(content={"status": "success", "message": f"Hello from FastAPI! You sent: {request_data.message}"})
|
22 |
+
|
23 |
+
# Ein optionaler Root-Endpunkt (oft gut für Health-Checks)
|
24 |
+
@app.get("/")
|
25 |
+
async def read_root():
|
26 |
+
"""
|
27 |
+
Root-Endpunkt für grundlegenden Health-Check.
|
28 |
+
"""
|
29 |
+
return {"message": "FastAPI is running!"}
|
30 |
+
|
31 |
+
print("INFO: FastAPI application script finished execution and defined 'app' variable.")
|
32 |
+
|
33 |
+
# Dieser Block wird in Hugging Face Spaces nicht direkt ausgeführt, da Uvicorn
|
34 |
+
# die App direkt lädt, aber er ist für lokale Tests nützlich.
|
35 |
+
if __name__ == "__main__":
|
36 |
+
import uvicorn
|
37 |
+
uvicorn.run(app, host="0.0.0.0", port=7860) # Lokaler Port 7860, wie in Space
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|